Paper deep dive
Lightweight GenAI for Network Traffic Synthesis: Fidelity, Augmentation, and Classification
Giampaolo Bovenzi, Domenico Ciuonzo, Jonatan Krolikowski, Antonio Montieri, Alfredo Nascita, Antonio Pescapè, Dario Rossi
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 3/27/2026, 1:37:30 AM
Summary
This paper introduces a lightweight Generative Artificial Intelligence (GenAI) pipeline for Network Traffic Generation (NTG), utilizing Transformer-based, state-space (SSM), and diffusion models to address data scarcity and privacy in Network Traffic Classification (NTC). By focusing on compact traffic representations (payload lengths and directions) rather than raw bytes, the proposed models achieve high fidelity and computational efficiency, with Transformers and SSMs (e.g., LLaMA, Mamba) outperforming traditional baselines in synthetic-only training and data augmentation tasks.
Entities (6)
Relation Signals (3)
Mamba → performson → Mirage-2019
confidence 95% · Experiments on two public datasets—Mirage-2019... demonstrate that lightweight GenAI models achieve strong fidelity
NetDiff-NR → usesarchitecture → Diffusion Models
confidence 95% · We use NetDiff-NR, a refined architecture operating on 2D Grammian Angular Summation Field (GASF) images.
LLaMA → improves → Network Traffic Classification
confidence 90% · Classifiers trained solely on synthetic traffic achieve up to 87% F1-score on real data.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Accurate Network Traffic Classification (NTC) is increasingly constrained by limited labeled data and strict privacy requirements. While Network Traffic Generation (NTG) provides an effective means to mitigate data scarcity, conventional generative methods struggle to model the complex temporal dynamics of modern traffic or/and often incur significant computational cost. In this article, we address the NTG task using lightweight Generative Artificial Intelligence (GenAI) architectures, including transformer-based, state-space, and diffusion models designed for practical deployment. We conduct a systematic evaluation along four axes: (i) (synthetic) traffic fidelity, (ii) synthetic-only training, (iii) data augmentation under low-data regimes, and (iv) computational efficiency. Experiments on two heterogeneous datasets show that lightweight GenAI models preserve both static and temporal traffic characteristics, with transformer and state-space models closely matching real distributions across a complete set of fidelity metrics. Classifiers trained solely on synthetic traffic achieve up to 87% F1-score on real data. In low-data settings, GenAI-driven augmentation improves NTC performance by up to +40%, substantially reducing the gap with full-data training. Overall, transformer-based models provide the best trade-off between fidelity and efficiency, enabling high-quality, privacy-aware traffic synthesis with modest computational overhead.
Tags
Links
- Source: https://arxiv.org/abs/2603.25507v1
- Canonical: https://arxiv.org/abs/2603.25507v1
Trouble viewing inline? Open PDF directly →
Full Text
46,296 characters extracted from source content.
Expand or collapse full text
Lightweight GenAI for Network Traffic Synthesis: Fidelity, Augmentation, and Classification Giampaolo Bovenzi, Domenico Ciuonzo, Jonatan Krolikowski, Antonio Montieri, Alfredo Nascita, Antonio Pescapè, Dario Rossi G. Bovenzi, D. Ciuonzo, A. Montieri, A. Nascita, and A. Pescapè are with the DIETI, University of Naples Federico I, Italy. J. Krolikowski and D. Rossi are with Huawei Technologies France SASU. Abstract Accurate Network Traffic Classification (NTC) is increasingly constrained by limited labeled data and strict privacy requirements. While Network Traffic Generation (NTG) provides an effective means to mitigate data scarcity, conventional generative methods struggle to model the complex temporal dynamics of modern traffic or/and often incur significant computational cost. In this article, we address the NTG task using lightweight Generative Artificial Intelligence (GenAI) architectures, including transformer-based, state-space, and diffusion models designed for practical deployment. We conduct a systematic evaluation along four axes: (i) (synthetic) traffic fidelity, (ii) synthetic-only training, (iii) data augmentation under low-data regimes, and (iviv) computational efficiency. Experiments on two heterogeneous datasets show that lightweight GenAI models preserve both static and temporal traffic characteristics, with transformer and state-space models closely matching real distributions across a complete set of fidelity metrics. Classifiers trained solely on synthetic traffic achieve up to 87%87\% F1-score on real data. In low-data settings, GenAI-driven augmentation improves NTC performance by up to +40%+40\%, substantially reducing the gap with full-data training. Overall, transformer-based models provide the best trade-off between fidelity and efficiency, enabling high-quality, privacy-aware traffic synthesis with modest computational overhead. Introduction Figure 1: Overview of the proposed lightweight GenAI-based network traffic generation pipeline and real-world challenges linked with our Research Questions. Training phase workflow: real network traces are segmented into biflows and mapped into canonical image- or token-based representations; these serve as inputs to train diffusion, transformer, or state-space generative models. Generation phase workflow: trained GenAI models are conditioned to generate image- or token-based representations for a given traffic class; these are then converted into traffic matrices to feed downstream network traffic classifiers; generation efficacy is assessed in terms of both model and traffic evaluation. The rapid evolution of Generative Artificial Intelligence (GenAI) is reshaping communications, accelerating the vision of autonomous and self-evolving networks [1]. Large generative models have demonstrated remarkable potential across telecom domains, exploiting massive datasets to learn complex patterns and generate new content. Applications already span physical-layer optimization and channel modeling [2], semantic communications, and network planning [1]. While their adoption for data-centric network monitoring and management [3] is also blooming, it remains an open frontier. Within this landscape, Network Traffic Classification (NTC) remains pivotal for network management and security across mobile, wired, and capillary Internet of Things (IoT) environments. While accurate NTC enables various critical functions, such as anomaly detection, quality of service provisioning, and security policy enforcement, traditional data-driven solutions often struggle to meet practical desiderata. Despite the advancements fueled by Machine Learning (ML), practical constraints like limited labeled data, imbalanced classes, and strict privacy regulations act as severe roadblocks. These constraints hinder scalability and reduce robustness in real-world settings [4]. To address these limitations, Network Traffic Generation (NTG) has been widely explored to synthesize realistic traffic data, enrich training sets, and strengthen classification robustness. Nevertheless, conventional approaches fall short. Classical data augmentation and traditional ML-based NTC rely on handcrafted statistical features, limiting adaptability to evolving traffic patterns and failing to capture rich temporal dependencies. Meanwhile, conventional NTG solutions struggle with a core trade-off: producing high-fidelity synthetic traffic while maintaining the computational efficiency required by modern networks [3]. In this article, we advocate for a shift toward lightweight GenAI architectures for practical NTG. As depicted in Fig. 1, we propose a modular NTG pipeline specifically tailored for GenAI. Rather than synthesizing fine-grained payload bytes using massive foundation models, our approach generates a compact traffic representation derived from header fields of the first packets in each network flow. This choice enables training and operating advanced GenAI models, namely Transformers, State Space Models (SSMs), and Diffusion Models (DMs), under a budget of 11–22 million parameters, ensuring computational efficiency without sacrificing generation fidelity. To validate this lightweight paradigm, we formulate four Research Questions (RQs) aligned with real-world networking challenges (Fig. 1, bottom left). We assess both generated data quality from different viewpoints (Traffic Evaluation – RQ 1 to RQ 3) and GenAI model efficiency (Model Evaluation – RQ 4): RQ 1: Can lightweight GenAI faithfully reproduce real traffic patterns? RQ 2: Can GenAI synthetic traffic enable privacy-preserving NTC without degrading performance? RQ 3: Can GenAI synthetic data mitigate training scarcity in low-data regimes? RQ 4: Are lightweight GenAI models computationally efficient for deployment? To answer these RQs, our work provides an extensive evaluation of synthetic traffic generated via lightweight GenAI models by (i) defining a comprehensive fidelity assessment procedure, and tackling two pivotal downstream NTC tasks, namely (ii) synthetic-only training and (iii) data augmentation for low-data regimes. Complementarily, we deliver an (iviv) usability assessment in terms of space and time complexity. Experiments on two public datasets— Mirage- 2019 (4040 mobile apps) and CESNET- TLS22- 80 (8080 network services)—demonstrate that lightweight GenAI models achieve strong fidelity, preserving both static and temporal traffic patterns. Compared to traditional NTG baselines (e.g., C V A E, SMOTE, and domain-expert transformations) and DMs, Transformers and SSMs (i.e. L L a M A and M a m b a, respectively) exhibit higher performance. Classifiers trained exclusively on synthetic data reach up to 87%87\% F1-score on real traffic, while GenAI-driven augmentation improves F1-scores by up to +40%+40\% in low-data regimes. Last but not least, resource analysis reveals a clear trade-off between architectural complexity and computational cost, with Transformer-based models offering the most favorable balance for practical deployment, combining a moderate memory footprint with the lowest generation latency. The remainder of the article explores the background and related work shaping the current NTG landscape and presents the proposed lightweight GenAI-based pipeline. It then details the experimental setup and discusses the results addressing the four RQs, before concluding with directions for future work. The GenAI Paradigm Shift in Network Traffic Generation This section frames NTG evolution, categorizing existing work into three phases: early statistical methods and conventional ML models, current large-scale GenAI models, and the emerging need for lightweight solutions addressing our research gap. Traditional Generative Methods. Early NTG relied on statistical generative models (e.g., Markov models) to capture sequential dependencies through state transitions. Though intuitive and lightweight, they do not scale well to mid- or long-range dependencies. Conventional ML models improved expressive power [5]. Variational Autoencoders learn probabilistic latent representations that enhance reconstruction accuracy, while Normalizing Flows provide exact likelihood estimation at the cost of a higher computational overhead. Generative Adversarial Networks, in contrast, focus on high-fidelity sampling but often suffer from training instability and mode collapse. Although traditional methods support downstream tasks like intrusion detection and traffic classification [3, 5], they struggle to balance training stability, computational efficiency, and fidelity in modeling complex, evolving traffic patterns. Recent GenAI Advancements. Since 2021, NTG has been impacted by the growing popularity of Transformers, DMs, and SSMs, each employing distinct strategies for traffic representation. Originating from Natural Language Processing (NLP), Transformer-based models (e.g., G P T s and T 5) [6, 7, 8, 9, 10] and SSMs (i.e. M a m b a) [11] treat network traffic as token sequences, learning networking “grammar” to model flow dynamics. Input sequences range from raw packet bytes [8, 11] and header fields (e.g., packet sizes, inter-arrival times, and directions [6]) to tcpdump/tshark packet summaries [7, 10]. To further bridge the modality gap between natural language and network data, the authors of [9] combine specialized traffic-domain tokenization and multimodal learning to understand expert instructions and learn task-specific traffic representations simultaneously. Crucially, generation strategies operate either iteratively, constructing flows “token-by-token” akin to sentences in NLP [8], or natively synthesize complete sequences [6, 11]. A notable exception is the work in [7], which generates Python code interacting with the Scapy library, acting more as traffic replay than GenAI synthesis. Conversely, DMs, originating from computer vision, have recently gained traction across the entire networking stack, from physical-layer channel generation and resource management [2] to NTG [12, 13]. DMs synthesize traffic by iteratively denoising random data until structured patterns emerge. This requires encoding traffic into image-like representations [12] or directly modeling raw byte streams [13] to capture high-fidelity details. Although these “large” models achieve high fidelity, their byte-level processing and iterative generation often make them impractical for network deployment. Positioning. Our work diverges from the emerging trend of massive network foundation models [10] to investigate the feasibility of lightweight GenAI. Indeed, while foundation models offer generalization and fine-tuning capabilities, they impose prohibitive computational costs. Instead, we explore NTG solutions based on Transformers, SSMs, and DMs under strict resource constraints (i.e. 11–22M parameters), prioritizing deployability and training/inference speed. Unlike related works focused on generating raw packet bytes [13, 11] (which increases complexity and risks payload data leakage), we synthesize lightweight traffic features constituting the network fingerprint of applications. Specifically, we model the time series of payload lengths and packet directions (× PL× DIR), akin to [8, 6, 12], but explicitly discard inter-arrival times, as these depend on network conditions rather than application logic [14]. Furthermore, we address the lack of rigorous validation in prior studies by introducing advanced fidelity metrics, including × PL× DIR n-grams and Markov transition matrices, to assess temporal integrity. Finally, targeting practical downstream tasks for NTG, our comprehensive evaluation demonstrates that lightweight GenAI enables effective classifier training even in (a) synthetic-only or (b) low-data regimes, ensuring efficiency for real-world deployment. Lightweight GenAI-Based NTG Pipeline at Work Figure 1 depicts the modular NTG pipeline powered by lightweight GenAI models, structured around two distinct phases. The training phase pre-processes real network traces to train GenAI models, while the generation phase leverages them to synthesize high-fidelity, application-conditioned traffic data. Training Phase Workflow. The training phase ingests real network traffic traces (e.g., from mobile apps or network services). To bypass the computational overhead and privacy risks associated with raw payload utilization, data undergoes traffic segmentation to group packets into bidirectional flows (biflows)111A biflow is a network flow consisting of packets flowing bidirectionally between the same network and transport endpoints—IPs, ports, and L4 protocol. It represents both directions of communication as a single entity., followed by feature extraction. This produces a highly efficient traffic matrix representation, where rows are packets and columns are packet fields. Specifically, we extract the Payload Length (PL) and Packet Direction (DIR) of the first 1010 packets of each biflow. Depending on the GenAI model, this matrix undergoes a specific modality-mapping before training: • TrafficMatrix2Image: Image-based models, such as DMs, interpret the (potentially preprocessed) traffic matrix as a structured 2D image. This enables DMs to learn complex 2D patterns that encode both the underlying features and the temporal dynamics of the flows, generating the entire traffic representation in a single, non-autoregressive step. • TrafficMatrix2Token: Sequence-based models, such as Transformers and SSMs, treat the matrix as a multivariate time series, serializing it into a sequence of discrete tokens, where each token encodes the vectorial values of the sequence steps. This enables autoregressive generation that explicitly captures complex temporal dependencies across packets and fields. Given this mapped data, the goal of the selected GenAI models is to learn a distribution that faithfully captures the underlying structure of the real network traces. Generation Phase Workflow. Once the GenAI models are trained, they are deployed to synthesize new traffic samples. Generation is conditioned via a <CLASS> token prompt, which dictates the target network class (e.g., a certain mobile app like YouTube, or a network service). The GenAI architectures generate synthetic samples in their native formats: 2D images for DMs or token sequences for Transformers and SSMs. To leverage these outputs for downstream NTC, an inverse mapping step is needed. Specifically, the Image2TrafficMatrix and Token2TrafficMatrix steps reconstruct the generated samples back into the original traffic matrix format, recovering the corresponding synthetic PLs and DIRs. Generation efficacy is evaluated from two complementary viewpoints: • Traffic Evaluation: To assess generation fidelity, we quantify the divergence between real and synthetic traffic distributions using distance metrics, such as the Jensen-Shannon Divergence (JSD). More precisely, we compute these distances across the traffic properties illustrated in Fig. 1: packet count histograms for session-level behavior, 1-gram histograms for marginal probabilities of PL and DIR, 2-gram histograms for temporal dependencies across consecutive packet pairs, and Markov transition matrices for first-order transition dynamics. Beyond fidelity, we evaluate the utility of generated traffic in two practical downstream NTC scenarios: synthetic-only training, where classifiers are trained exclusively on synthetic data and tested on real samples, and data augmentation, assessing whether enriching a few real samples with synthetic ones boosts classification performance. • Model Evaluation: To validate the deployment feasibility of our lightweight GenAI models, we also profile their computational efficiency during training and inference, evaluating training time, generation latency, GPU memory utilization, and on-disk model footprint. Additionally, we investigate post-training quantization to assess whether these architectures can be further optimized for resource-constrained environments. Lightweight GenAI Models. To implement our NTG pipeline under strict computational constraints, we leverage lightweight GenAI models (≈1–2≈\!1--2M parameters) belonging to different families: • Diffusion Models (DMs): DMs iteratively reverse a noising process to reconstruct realistic samples. We use - N e t D i f f u s- N R [12], a refined architecture operating on 2D Grammian Angular Summation Field (GASF) images. A post-generation refinement step accurately maps the 2D GASF images back to traffic sequences, minimizing reconstruction errors and improving the quality of the synthesized traffic traces.222We exclude heavier byte-level alternatives like N e t D i f f u s i o n [13] from our evaluation, as its massive scale (hundreds of millions of parameters) and simplified direction modeling contradict our lightweight, time-series focus. • Transformer-based Models: Transformers excel at sequence modeling by capturing temporal dependencies through self-attention. We adopt L L a M A [15], a causal decoder-only architecture designed to model long-range dependencies efficiently. It leverages optimized attention mechanisms to autoregressively synthesize high-fidelity sequences while maintaining computational scalability. • Structured State-Space Models (SSMs): To complement Transformers, we explore SSMs designed for efficient sequence processing. We employ M a m b a [11], which replaces standard attention with a selective state-space formulation. Operating causally, it achieves linear-time scalability, well-suited for modeling per-biflow traffic sequences. Experimental Evaluation This section evaluates our lightweight GenAI-based NTG pipeline to answer the four RQs formulated in the Introduction. First, we outline the experimental setup, encompassing the employed datasets and generation configurations. Then, we provide the corresponding Research Answers (RAs), covering the traffic evaluation for generation fidelity (RA 1) and downstream NTC utility (RA 2 and RA 3), followed by the model evaluation profiling computational efficiency and deployment feasibility (RA 4). TABLE I: Configuration of the GenAI models, grouped by main attributes. The last column provides the repository link. Model HS IS #L #AH Repo C V A E 500/250500/250 2020 66 – – L L a M A 160160 320320 44 88 M a m b a 7272 144144 44 – - N e t D i f f u s- N R 3232 – 44 44 Legend: HS – Hidden Size; IS – Intermediate Size; #L – Number of Layers; #AH – Number of Attention Heads; Repo – Repository (clickable icon). Experimental Setup. Our evaluation relies on two public network traffic datasets: Mirage- 2019 333https://traffic.comics.unina.it/mirage/mirage-2019.html, containing 4040 Android apps with ≈100≈\!100k biflows, and CESNET- TLS22- 80, a CESNET- TLS22 subset444https://w.liberouter.org/datasets/cesnet-tls22 downsampled to cover the top 8080 services and obtain a sample size comparable to Mirage- 2019. Raw traffic data are pre-processed into sequences of the first 1010 signed PLs (± ), where negative and positive values encode downstream and upstream DIRs, respectively. For a fair cross-architecture comparison, all lightweight GenAI models are bounded to 11–22M trainable parameters. Alongside these models, we employ a Conditional Variational Autoencoder ( C V A E) baseline for class-aware traffic generation. Table I summarizes their key hyperparameters. For sequence-based models, the vocabulary assigns a unique token to each signed PL value, resulting in 2×PLmax2×PL_ possible tokens, augmented with an <EOS> token and NN <CLASS> tokens (N=40N=40 for Mirage- 2019 and N=80N=80 for CESNET- TLS22- 80). To ensure fixed-length inputs during training, biflows with fewer than 1010 packets are right-padded via <PAD> tokens. (a) Mirage- 2019 (b) CESNET- TLS22- 80 Figure 2: Radar plots of 66 fidelity metrics comparing real and synthetic traffic data across generative models for Mirage- 2019 (left) and CESNET- TLS22- 80 (right). For all considered metrics, lower values indicate better performance (with 0 being optimal). Note that the axes are scaled with 0 at the outer edge, meaning that models producing larger polygon areas exhibit higher generative fidelity. RA1 – Fidelity Evaluation. To answer RQ 1, we quantitatively assess the fidelity of generated traffic through the six metrics reported in the radar plots of Fig. 2. First, we translate the visual properties depicted in Fig. 1 into numerical metrics by computing the macro-averaged JSD between real and synthetic distributions for packet count histograms (JSDNumPacketsJSD_NumPackets), 1-gram histograms (JSD1-gramJSD_1-gram), 2-gram histograms (JSD2-gramJSD_2-gram), and Markov transition matrices (JSDMarkovJSD_Markov). Furthermore, we assess the realism and privacy of generated biflows via two additional metrics [5]: (i) UniqAlignevaluates data realism by computing the uniqueness score (i.e. the proportion of distinct sequences) independently for the real and synthetic datasets, and then measuring their absolute difference; a lower score indicates that synthetic traffic accurately replicates the repetition patterns of real data. (i) Leakagedirectly quantifies the exact sequence overlap between real and synthetic datasets via Jaccard similarity; a lower score indicates novel biflow generation rather than mere memorization of training data, thereby mitigating privacy leakage risks. Figure 2 summarizes the performance of generative models. Since all metrics follow a “lower-is-better” logic, radar plot axes are inverted, meaning that larger areas correspond to higher generation fidelity. Results are consistent across Mirage- 2019 and CESNET- TLS22- 80. L L a M A and M a m b a outperform all alternatives, achieving near-zero JSD scores across all evaluated traffic properties. They effectively capture both marginal × PL× DIR distributions (11-gram) and more complex temporal dependencies (22-gram and Markov). Notably, L L a M A achieves the best JSDMarkovJSD_Markov, confirming its ability to model advanced sequential transitions. Also, both sequence-based models exhibit near-optimal UniqAlign and Leakage scores. This demonstrates that their generated samples closely match the real traffic distribution, synthesizing highly diverse and realistic sequences without merely memorizing the training set (with M a m b a showing a slight edge over L L a M A in leakage mitigation). Interestingly, the C V A E baseline suitably performs only on coarse-grained properties, successfully matching the biflow length distribution (JSDNumPackets≈0JSD_NumPackets\!≈\!0). However, its fidelity drops when capturing complex patterns, exposing its structural limitations in modeling fine-grained traffic dynamics, though it still limits data leakage. - N e t D i f f u s- N R consistently yields the lowest fidelity. It struggles to capture structural traffic properties, exhibiting the highest JSD scores (i.e. the smallest polygon area) and failing to accurately model even the biflow lengths. Despite these limitations, it achieves a moderate UniqAlign and successfully minimizes Leakage. TABLE I: F1-scores of an RF classifier trained on synthetic and tested on real Mirage- 2019 (orange) and CESNET- TLS22- 80 (azure) traffic. Best GenAI models are in bold. (GenAI) Approach Mirage- 2019 CESNET- TLS22- 80 Train on Real 85.84%85.84\% 91.80%91.80\% C V A E 66.73%66.73\% 75.44%75.44\% - N e t D i f f u s- N R 46.83%46.83\% 65.84%65.84\% L L a M A 78.78% 87.43% M a m b a 76.07%76.07\% 85.09%85.09\% RA2 – NTC: Train on Synthetic Traffic. To address RQ 2, we evaluate generated data utility via a train-on-synthetic, test-on-real approach. We train a Random Forest (RF) downstream classifier exclusively on synthetic samples and evaluate its generalization on unseen real traffic from Mirage- 2019 and CESNET- TLS22- 80. Performance under this setting reflects both the fidelity of the synthetic samples and their alignment with real-world class distributions. Table I shows that RF models trained on synthetic traffic exhibit an expected performance gap relative to real-data training, reflecting the inherent difficulty of fully reproducing realistic traffic characteristics. Nevertheless, L L a M A and M a m b a generated data lead to consistently higher classification performance, outperforming the C V A E baseline and - N e t D i f f u s- N R. On Mirage- 2019, the RF trained on L L a M A and M a m b a samples achieves 78.78%78.78\% and 76.07%76.07\% F1-scores, respectively, substantially reducing the gap with the real-data upper bound (85.84%85.84\%). Similarly, on CESNET- TLS22- 80, the synthetic-trained RF reaches 87.43%87.43\% ( L L a M A) and 85.09%85.09\% ( M a m b a) F1-scores, closely trailing the 91.80%91.80\% obtained with real traffic. In contrast, - N e t D i f f u s- N R severely underperforms across both datasets, confirming that this DM struggles to capture traffic characteristics relevant for downstream classification. To summarize, L L a M A and M a m b a offer the best balance of generation fidelity and utility for downstream tasks, enabling synthetic-to-real generalization and privacy-preserving NTC with minimal performance degradation, whereas - N e t D i f f u s- N R appears less suitable for realistic traffic synthesis. (a) Mirage- 2019 (b) CESNET- TLS22- 80 Figure 3: F1-score in data augmentation scenarios under low-data regimes for Mirage- 2019 (left) and CESNET- TLS22- 80 (right) using an RF classifier. Colors indicate the approach family: orange for sequence-based GenAI, green for other generative models, red for statistical techniques, violet for expert transformations, black for real-only training. RA3 – NTC: Data Augmentation. We investigate GenAI-driven data augmentation under data-scarcity conditions, where only a limited fraction of real training samples is available to the downstream classifier. The GenAI model is trained on the full labeled dataset and used to generate synthetic samples. The downstream classifier is trained with few real samples plus synthetic data. This reflects a practical use case of a network operator with limited data leveraging a pre-trained GenAI model to augment the training dataset. To establish a more extensive benchmark, we compare GenAI models against two non-AI baselines: (i) Fast Retransmit [14], a domain-expert traffic transformation, and (ii) SMOTE, a statistical synthesis technique. Fast Retransmit probabilistically delays a single packet to mimic a TCP retransmission, while SMOTE generates new samples by replicating or interpolating existing ones. Unlike GenAI models, which are trained offline on the entire dataset, Fast Retransmit and SMOTE operate directly on the limited data available at augmentation time. In detail, we evaluate augmentation utility by training an RF classifier on mixed real and synthetic traffic. To this end, we upsample all classes with synthetic samples to match the size of the majority class from the original training set, yielding a perfectly balanced set. In the low-data regime (i.e. 55–20%20\% of the original training set), Fig. 3 shows that GenAI-based augmentation produces substantial F1-score improvements. On Mirage- 2019, L L a M A and M a m b a samples significantly boost the classification performance compared to real-only training, rapidly approaching the F1-scores achieved with abundant real data. On CESNET- TLS22- 80, the gains remain consistent, albeit more moderate (1010–15%15\% F1-score improvement over training without augmentation). Conversely, traditional statistical augmentation methods exhibit limited impact: SMOTE typically matches, or even slightly underperforms, real-only training. Fast Retransmit exhibits more favorable behavior, yielding moderate gains in the low-data regime on CESNET- TLS22- 80. Nonetheless, its impact diminishes on Mirage- 2019 and remains strictly below that of L L a M A and M a m b a. The other generative methods exhibit markedly different behaviors across datasets. C V A E consistently improves upon real-only training, confirming its ability to model relevant traffic characteristics. - N e t D i f f u s- N R, instead, shows limited effectiveness, yielding marginal gains on Mirage- 2019 and systematically underperforming on CESNET- TLS22- 80. Overall, sequence-based GenAI augmentation consistently achieves the best performance across both datasets, highlighting its clear superiority over traditional statistical, expert-driven, and other generative alternatives. TABLE I: Training and generation resource usage (time, GPU utilization, memory, and on-disk size) for each model on a high-end datacenter GPU (48GB), measured over 10 runs and reported as medians. Training: 1010 epochs, 55 classes, 500500 samples/class, batch size 11. Generation: 100100 samples/class, batch size 11. Model Training Generation Model Size Time [[s/epoch]] GPU [%][\%] Mem [[MB]] Time [[ms/sample]] GPU [%][\%] Mem [[MB]] on Disk [[MB]] CVAE 22.568 022.568 1818 379 0379 0.50 000.50 1 01 291 0291 3.9 03.9 - N e t D i f f u s- N R 117.029117.029 2121 387 0387 860.67860.67 1919 355 0355 4.2 04.2 L L a M A 36.810 036.810 2020 393 0393 31.21 031.21 1717 353 0353 7.9 07.9 M a m b a 108.477108.477 1515 377 0377 148.52148.52 1414 359 0359 15.515.5 L L a M A-- PTQ_ int8- WO — — — 45.33 045.33 1616 357 0357 3.5 03.5 L L a M A-- PTQ_ int8- DA — — — 990.04990.04 1414 347 0347 3.4 03.4 RA4 – Computational Efficiency. Addressing RQ 4, we analyze computational efficiency by measuring training time, generation latency, GPU utilization, memory consumption, and on-disk footprint. All generative models were trained for 1010 epochs with 500500 samples per class (55 classes, batch size of 11) using a high-end datacenter GPU (4848 GB), and subsequently used to generate 100100 synthetic samples per class. The results reveal a clear trade-off between computational cost and architectural complexity. During training, times range from 22.622.6 s/epoch for C V A E to ≈108≈\!108–117117 s/epoch for models with more structured generative mechanisms. Notably, L L a M A stands out as an exception, achieving a highly competitive 36.836.8 s/epoch despite its autoregressive architecture. - N e t D i f f u s- N R exhibits the highest GPU utilization during training (21%21\%), followed by L L a M A (20%20\%) and C V A E (18%18\%), whereas M a m b a is the most efficient (15%15\%). Memory usage remains uniform across all models, ranging from 377377 MB ( M a m b a) to 393393 MB ( L L a M A), suggesting that the dominant memory cost is the framework overhead rather than the model itself. Generation latency varies substantially. C V A E is the fastest (0.500.50 ms/sample), followed by L L a M A (31.2131.21 ms/sample), whereas M a m b a and - N e t D i f f u s- N R incur considerably higher latencies (148.52148.52 ms/sample and 860.67860.67 ms/sample, respectively). Regarding on-disk footprint, M a m b a is by far the largest model at 15.515.5 MB, while the others range from 3.93.9 MB ( C V A E) to 7.97.9 MB ( L L a M A). Lastly, we explore Post-Training Quantization (PTQ) to assess whether these models can be further optimized for resource-constrained environments. We apply PTQ exclusively to L L a M A, as its optimal trade-off between architectural footprint (substantially smaller than M a m b a) and generation fidelity makes it the ideal candidate for a tiny-footprint generative architecture. Specifically, we investigate two int8 PTQ variants: weight-only ( L L a M A-- PTQ_ int8- WO) and dynamic activation ( L L a M A-- PTQ_ int8- DA). The weight-only variant preserves GPU memory consumption while more than halving the model size (from 7.97.9 MB to 3.53.5 MB) and reducing GPU utilization from 19%19\% to 16%16\%. This enhancement comes at the cost of a moderate increase in generation latency (45.3345.33 ms/sample). Conversely, the dynamic activation variant further reduces the model size (3.43.4 MB) and GPU utilization (14%14\%), but incurs a substantial latency penalty (≈+960≈\!+960\,ms/sample) due to the overhead of on-the-fly activation quantization. Notably, both quantized L L a M A variants achieve fidelity metrics and real-traffic classification performance consistent with the non-quantized model up to two significant figures, confirming that quantization introduces no meaningful degradation. Taken together, these findings demonstrate that lightweight GenAI architectures, particularly L L a M A coupled with quantization, can achieve the computational efficiency required for practical deployment. Conclusion This work presented a comprehensive study on generative approaches for network traffic synthesis, focusing on fidelity, downstream classification, and deployment feasibility. Overall, our results highlight L L a M A and M a m b a as the most promising models for realistic traffic synthesis, privacy-preserving classification, and effective data augmentation. In fidelity assessment (RQ 1), both models achieve near-zero JSD scores across all properties; L L a M A excels on Markov transition matrices, while M a m b a shows a slight edge in leakage mitigation. In synthetic-only training (RQ 2), L L a M A and M a m b a reach 78.78%78.78\% and 76.07%76.07\% F1-score on Mirage- 2019, and 87.43%87.43\% and 85.09%85.09\% on CESNET- TLS22- 80, narrowing the gap with real-data training to ≈9%≈\!9\% and ≈13%≈\!13\%, respectively. In data augmentation (RQ 3), sequence-based GenAI improves classification by up to +40%+40\% F1-score in the low-data regime (5–20%5--20\% of real training set) on Mirage- 2019, with consistent +10–15%+10--15\% gains on CESNET- TLS22- 80. Regarding computational efficiency (RQ 4), L L a M A offers the best trade-off with 36.836.8\,s/epoch training time, 31.2131.21\,ms/sample generation latency, and a 7.97.9\,MB on-disk footprint, further reducible to 3.53.5\,MB via int8 weight-only post-training quantization, without meaningful degradation in generation fidelity or classification performance. Conversely, diffusion models (- N e t D i f f u s- N R) and baselines ( C V A E) proved less effective, either incurring prohibitive generation latencies or failing to capture fine-grained temporal dynamics. Future work will explore adaptive generation strategies and hybrid pipelines combining generative models with domain-specific transformations. Furthermore, we aim to evaluate these lightweight architectures in real-world deployment scenarios, such as integrating quantized models into edge-based intrusion detection systems for on-the-fly, privacy-preserving traffic analysis. Lastly, extending this paradigm toward Agentic AI powered by lightweight models represents a promising frontier to enable autonomous, closed-loop network simulation and proactive defense mechanisms. References Bariah et al. [2024] L. Bariah et al., “Large generative AI models for telecom: The next big thing?” IEEE Commun. Mag., vol. 62, no. 11, p. 84–90, 2024. Xu et al. [2024] X. Xu et al., “Generative artificial intelligence for mobile communications: A diffusion model perspective,” IEEE Commun. Mag., 2024. Bovenzi et al. [2025] G. Bovenzi et al., “Mapping the landscape of generative AI in network monitoring and management,” IEEE Trans. Netw. Service Manag., 2025. Aceto et al. [2023] G. Aceto et al., “AI-powered internet traffic classification: Past, present, and future,” IEEE Commun. Mag., vol. 62, no. 9, p. 168–175, 2023. Aceto et al. [2024] —, “Synthetic and privacy-preserving traffic trace generation using generative ai models for training network intrusion detection systems,” Journal of Network and Computer Applications, p. 103926, 2024. Bikmukhamedov and Nadeev [2021] R. F. Bikmukhamedov and A. F. Nadeev, “Multi-class network traffic generators and classifiers based on neural networks,” in Systems of Signals Generating and Proc. in the Field of on Board Comm., 2021. Kholgh and Kostakos [2023] D. K. Kholgh and P. Kostakos, “PAC-GPT: A novel approach to generating synthetic network traffic with GPT-3,” IEEE Access, vol. 11, p. 114 936–114 951, 2023. Qu et al. [2024] J. Qu et al., “TrafficGPT: Breaking the token barrier for efficient long traffic analysis and generation,” arXiv preprint arXiv:2403.05822, 2024. Cui et al. [2025] T. Cui et al., “TrafficLLM: Enhancing large language models for network traffic analysis with generic traffic representation,” arXiv preprint arXiv:2504.04222, 2025. Mayhoub et al. [2026, in press] S. Mayhoub et al., “Talk like a packet: Rethinking network traffic analysis with transformer foundation models,” IEEE Commun. Mag., 2026, in press. Chu et al. [2024] A. Chu et al., “Feasibility of state space models for network traffic generation,” in Proc. of the SIGCOMM Workshop on Networks for AI Computing, 2024, p. 9–17. Sivaroopan et al. [2024] N. Sivaroopan et al., “NetDiffus: Network traffic generation by diffusion models through time-series imaging,” Computer Networks, vol. 251, p. 110616, 2024. Jiang et al. [2024] X. Jiang et al., “Netdiffusion: Network data augmentation through protocol-constrained traffic generation,” Proc. of the ACM on Measurement and Analysis of Computing Systems, vol. 8, no. 1, p. 1–32, 2024. Wang et al. [2024] C. Wang et al., “Data augmentation for traffic classification,” in Int. Conf. on Passive and Active Network Measurement, 2024, p. 159–186. Touvron et al. [2023] H. Touvron et al., “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023. Giampaolo Bovenzi (giampaolo.bovenzi@unina.it) is an Assistant Professor at the University of Napoli Federico I. His research concerns (anonymized and encrypted) traffic classification and network security. Domenico Ciuonzo [SM] (domenico.ciuonzo@unina.it) is an Associate Professor at the University of Napoli Federico I. His research concerns data fusion, network analytics, IoT, signal processing, and AI. Jonatan Krolikowski (jonatan.krolikowski@huawei.com) is a senior research engineer at the DataCom Lab of Huawei’s Paris Research Center. His research interests include ML- and operations research-driven optimization of real-world networks and the modeling and analysis of network-related problems, with a recent focus on time series modeling. Antonio Montieri (antonio.montieri@unina.it) is an Assistant Professor at the University of Napoli Federico I. His research concerns network measurements, traffic classification, modeling and prediction, and AI for networks. Alfredo Nascita (alfredo.nascita@unina.it) is an Assistant Professor at the University of Napoli Federico I. His research interests include Internet network traffic analysis, machine and deep learning, and explainable artificial intelligence. Antonio Pescapé [SM] (pescape@unina.it) is a Full Professor at the University of Napoli Federico I. His work focuses on measurement, monitoring, and analysis of the Internet. Dario Rossi [SM] (dario.rossi@huawei.com) is network AI CTO and director of the DataCom Lab at Huawei Technologies, France. He has coauthored 15+ patents and over 200+ papers in leading conferences and journals, and has received 9 best paper awards, a Google Faculty Research Award (2015), and an IRTF Applied Network Research Prize (2016).